OCR

A 0.9-billion-parameter OCR model just topped the OmniDocBench leaderboard, ahead of Gemini 3.1 Pro and GPT-5.4. For anyone processing documents at volume, that flips the architecture question.

Vision language models now read scanned title packages with accuracy that rivals dedicated OCR. Here's what that shift means for Quebec title review and the hybrid pipeline inside Cleardeal.